Improving the Effectiveness of Performance Analysis for HPC by Using Appropriate Modeling and Simulation Schemes

Performance modeling and simulation of parallel applications are critical performance analysis techniques in High Performance Computing (HPC). Efficient and accurate performance modeling and simulation can aid the tuning and optimization of current systems as well as the design of future HPC systems. As the HPC applications and systems increase in size, efficient and accurate performance modeling and simulation of parallel applications is becoming increasingly challenging. In general, simulation yields higher accuracy at the cost of high simulation time in comparison to modeling. This dissertation aims at developing effective performance analysis techniques for the next generation HPC systems. Since modeling is often orders of magnitude faster than simulation, the idea is to separate HPC applications into two types: 1) the ones that modeling can produce similar performance results as simulation and 2) the ones that simulation can result in more meaningful information about the application performance than modeling. By using modeling for the first type of applications and simulation for the rest of applications, the efficiency of performance analysis can be significantly improved. The contribution of this thesis is three-fold. First, a comprehensive study of the performance and accuracy trade-offs between modeling and simulation on a wide range of HPC applications is performed. The results indicate that for the majority of HPC applications, modeling and simulation yield similar performance results. This lays the foundation for improving performance analysis on HPC systems by selecting between modeling and simulation on each application. Second, a scalable and fast classification techniques (MFACT) are developed based on the Lamport's logical clock that can provide fast diagnosis of MPI application performance bottleneck and assist in the processing of application tuning and optimization on current and future HPC systems. MFACT also classifies HPC applications into bandwidth-bound, latency-bound, communication-bound, and computation-bound. Third, built-upon MFACT, for a given system configuration, statistical methods are introduced to classify HPC applications into the two types: the ones that needs simulation and the ones that modeling is sufficient. The classification techniques and tools enable effective performance analysis for future HPC systems and applications without losing accuracy.